Mingduo Zhao
(pronounced Ming-dwo Jow)

Abstract

AI systems for dynamic pricing, targeted promotions, and individualized recommendations typically assume that observed consumer behavior truthfully reveals underlying preferences. However, when consumers recognize that their actions influence future targeting decisions, behavior becomes strategic rather than preference-revealing, undermining the validity of standard machine learning based targeting. In order to address this challenge, we introduce Structural Transfer Learning (STL), a new framework that incorporates structural economic modeling into machine learning pipelines to account for strategic responses induced by policy interventions. STL constructs policy dependent in stance weights that adjust for endogenous domain shifts, enabling firms to learn targeting policies that remain effective even when consumers actively game the system. Furthermore, we demonstrate the practical value of STL through a stylized online experiment in a consumer research setting. While simplified, the design captures a core strategic response common to many personalization environments, where targeting rules shape behavior. In this setting, naive behavioral targeting rules lead to substantial misallocation of incentives. Applying STL improves expected profits by up to 35% relative to the naive machine learning benchmark that ignores strategic responses, depending on the relative cost of incentives. These findings highlight the need for a fundamental shift in firms’ personalization strategies-from solely optimizing predictive models to designing incentive mechanisms that are robust to strategic consumer behavior.

Abstract

In online marketplaces, consumers rely on reviews to make informed purchase decisions, making the presence of fake reviews detrimental. Previous literature implies that products with fake reviews can display some patterns in review distribution, such as a higher discrepancy in ratings. Consumers might take this pattern into account when making their purchase decisions. In this paper, we explore the interplay between fake reviews and ratings discrepancy, and their impact on consumer demand, while controlling for average product ratings. First, using a data set with fake review labels, we find that product ratings discrepancy is positively correlated with the probability that the product has fake reviews. Second, through an identification strategy exploiting ratings discrepancy changes due to rating distribution rounding, we find evidence consistent with a negative causal impact of ratings discrepancy on consumer demand. Then, we conduct two experiments to establish and quantify the mechanism of the impact of ratings discrepancy on consumer demand through consumer suspicion of fake reviews. The first experiment shows that higher ratings discrepancy increases consumer suspicion of fake reviews, and the second experiment shows that heightened suspicion reduces consumer willingness to pay. Together, these findings reveal that consumers use ratings discrepancies as a signal of fake reviews, and this suspicion impacts their purchase decisions. The findings highlight the importance of understanding the relationship between fake reviews, ratings discrepancies, and consumer demand in online marketplaces.

Abstract

Are individuals from marginalized groups perceived as less persuasive? Does this perception lead to their underrepresentation in visible, influential roles? Does it also contribute to their reluctance to share their views? These are important questions that require further investigation. Amazon reviews provide a valuable and quantifiable context to explore these questions. This paper investigates how the perceived identity of reviewers influences helpfulness votes, the ranking of their reviews, and their willingness to post new reviews on Amazon. Utilizing a data set of over 1.8 million reviews, we apply advanced natural language processing and computer vision tools to infer the gender and ethnicity of reviewers based on their user aliases and avatars. Combined with experimental data, our analysis shows causal evidence that reviews perceived as authored by women, non-binary individuals, or ethnic minorities are rated as less helpful than those associated with men or white individuals. Furthermore, we show that Amazon’s review ranking algorithm amplifies initial disparities in helpfulness recognition. Because reviews from marginalized groups receive fewer votes, they are ranked lower and receive less exposure, which in turn leads to even fewer votes-creating a self-reinforcing loop that perpetuates the underrepresentation of marginalized voices. We also find evidence of self-censorship: reviewers are less willing to continue contributing after receiving fewer helpfulness votes on their previous reviews. To address these disparities, we propose two platform-level interventions-identity blind review displays and affirmative action mechanisms within the review ranking algorithm. To evaluate their impact, we further estimate a structural model that captures both the supply and demand sides of the review ”market” to assess the welfare implications of these interventions.

Abstract

In recent years, numerous political candidates have achieved electoral success after first becoming well-known as influencers in non-political fields. Famous examples include Donald Trump, Ronald Reagan, Arnold Schwarzenegger, and others. This study explores the relationship between one’s fame as an influencer and their performance in U.S. gubernatorial elections from 1865 to 2020. With a dataset of 3,942 candidates and their Wikipedia biographies, we use Large Language Models (LLMs) to identify influencers as individuals who achieved public prominence through roles such as entrepreneurs, professional experts, or athletes before entering politics. The findings reveal that influencer candidates hold a significant and substantial advantage and in some cases are 20% more likely to win governor elections. This effect is particularly pronounced in three cases: Republican candidates in general, especially those running in swing states, and individuals making their first run for office and/or lacking prior experience as a governor. As is common in much of the existing literature in this field, establishing causality presents a common challenge. To address this, we conduct a voting experiment with approximately 1,800 participants. The results show that the estimated advantage of being an influencer closely aligns with findings from the observational analysis, reinforcing the causal interpretation. This research provides important insights into the dynamics of American democracy, indicating that public recognition and celebrity status can serve as pathways to electoral success.

Abstract

This study examines how media ownership consolidation influences the effectiveness of earned media in shaping political brands in U.S. mayoral elections. Using data from over 300 local radio markets between 1945 and 2006, we link radio station ownership records and news programming to local election outcomes. We find that increased market concentration strengthens the re-election prospects of incumbent mayors while hindering the performance of challengers. The effect of increased concentration is primarily driven by a reduction in the number of news-producing stations, and it dispro portionately benefits lower-quality incumbents. We theoretically model this phenomenon within an oligopoly framework, demonstrating how ownership structure affects news quality, and ultimately, electoral outcomes. To address potential endogeneity in market structure, we leverage a novel instrument based on multi-market mergers. From both managerial and policy perspectives, our results underscore the broader consequences of media consolidation: it not only weakens the informative performance of earned media but also lowers the quality of decision-making, which poses serious risks to the electoral process, as it may prevent the most qualified candidates from winning. Additionally, for firms, more concentrated media ownership reduces opportunities for visibility through earned media, diminishing the impact of superior offerings and limiting entry of new products.

Abstract

We propose Deep Longitudinal Targeted Minimum Loss-based Estimation (Deep LTMLE), a novel approach to estimate the counterfactual mean of outcome under dynamic treatment policies in longitudinal problem settings. Our approach utilizes a transformer architecture with heterogeneous type embedding trained using temporal-difference learning. After obtaining an initial estimate using the transformer, following the targeted minimum loss-based likelihood estimation (TMLE) framework, we statistically corrected for the bias commonly associated with machine learning algorithms. Furthermore, our method also facilitates statistical inference by enabling the provision of 95% confidence intervals grounded in asymptotic statistical theory. Simulation results demonstrate our method’s superior performance over existing approaches, particularly in complex, long time-horizon scenarios. It remains effective in small-sample, short-duration contexts, matching the performance of asymptotically efficient estimators. To demonstrate our method in practice, we applied our method to estimate counterfactual mean outcomes for standard versus intensive blood pressure management strategies in a real-world cardiovascular epidemiology cohort study.